a question about BG loss

[1173206772]

hello,your job is great。but I have some questions about the BGLOSS .
The first item is to make the prediction in the foreground area(distracting objects) near to 0, and the second one is to avoid learning a trivial solution (the paper says that we do not have the ground truth for the background area). however. can we directly use CELoss, like this nn.CELoss(output,1-label) to achieve BGLoss,？I think that in this way, I can force the foreground part to learn 0 and the background part to learn 1, so I don't need the second loss mentioned in the paper to avoid trivial solutions。
Did I understand wrong？thank you

[LIUYUANWEI98]

Using nn.CELoss(output, 1-label) poses issues. It fails to consider other potential foreground objects within the background region, treating them all as background. Over multiple iterations, this approach disrupts the learning process.